[i-g-t v3 2/4] lib/xe/xe_util: Introduce helper functions
Zbigniew Kempczyński
zbigniew.kempczynski at intel.com
Thu Feb 20 07:15:13 UTC 2025
On Tue, Feb 18, 2025 at 06:57:53PM -0500, Oak Zeng wrote:
> From: Bommu Krishnaiah <krishnaiah.bommu at intel.com>
>
> Introduce helper functions for buffer creation, binding,
> destruction and command submission etc. With those helpers,
> writing a xe igt test will be much easier, which will be
> showed in a coming example.
>
> v2: use to_user_pointer to cast a pointer (Kamil)
> s/insert_store/xe_insert_store (Kamil)
> s/cmdbuf_fill_func_t/xe_cmdbuf_fill_func_t (Kamil)
> v3: refactor command buffer fill interface (Zbigniew)
>
> Signed-off-by: Bommu Krishnaiah <krishnaiah.bommu at intel.com>
> Signed-off-by: Oak Zeng <oak.zeng at intel.com>
> Cc: Himal Prasad Ghimiray <himal.prasad.ghimiray at intel.com>
> ---
> lib/xe/xe_util.c | 239 +++++++++++++++++++++++++++++++++++++++++++++++
> lib/xe/xe_util.h | 47 ++++++++++
> 2 files changed, 286 insertions(+)
>
> diff --git a/lib/xe/xe_util.c b/lib/xe/xe_util.c
> index 06b378ce0..891316ecd 100644
> --- a/lib/xe/xe_util.c
> +++ b/lib/xe/xe_util.c
> @@ -13,6 +13,245 @@
> #include "xe/xe_query.h"
> #include "xe/xe_util.h"
>
> +/**
> + * xe_cmdbuf_exec_ufence_gpuva:
> + * @cmdbuf Pointer to the xe_cmdbuf structure representing the command buffer.
> + *
> + * Returns the GPU virtual address of the execution user fence located at the
> + * end of the command buffer.
> + */
> +static uint64_t xe_cmdbuf_exec_ufence_gpuva(struct xe_cmdbuf *cmdbuf)
> +{
> + /* the last 8 bytes of the cmd buffer is used as ufence */
> + return cmdbuf->buf.gpu_addr + cmdbuf->buf.size - UFENCE_LENGTH;
I think instead of hardcoding 8 as length of the fence you should use
sizeof() of timeline_value from drm_xe_sync.
> +}
> +
> +/**
> + * xe_cmdbuf_exec_ufence_cpuva:
> + * @cmdbuf Pointer to the xe_cmdbuf structure representing the command buffer.
> + *
> + * Returns the CPU virtual address of the execution user fence located at the
> + * end of the command buffer.
> + */
> +static void *xe_cmdbuf_exec_ufence_cpuva(struct xe_cmdbuf *cmdbuf)
> +{
> + /* the last 8 bytes of the cmd buffer is used as ufence */
> + return (char *)cmdbuf->buf.cpu_addr + cmdbuf->buf.size - UFENCE_LENGTH;
> +}
> +
> +
> +/**
> + * __xe_submit_cmd:
> + * @cmdbuf Pointer to the command buffer structure
> + *
> + * Submits a command buffer to the GPU, waits for its completion, and verifies
> + * the user fence value
> + *
> + * Return: The result of waiting for the user fence value
> + */
> +int64_t __xe_submit_cmd(struct xe_cmdbuf *cmdbuf)
> +{
> + int64_t timeout = NSEC_PER_SEC;
> + int ret;
> +
You should assert if cmdbuf is null.
> + struct drm_xe_sync sync[1] = {
> + { .type = DRM_XE_SYNC_TYPE_USER_FENCE,
> + .flags = DRM_XE_SYNC_FLAG_SIGNAL,
> + .timeline_value = USER_FENCE_VALUE,
> + .addr = xe_cmdbuf_exec_ufence_gpuva(cmdbuf),},
> + };
> + struct drm_xe_exec exec = {
> + .num_batch_buffer = 1,
> + .num_syncs = 1,
> + .syncs = to_user_pointer(sync),
> + .exec_queue_id = cmdbuf->exec_queue,
> + .address = cmdbuf->buf.gpu_addr,
> + };
> +
> + xe_exec(cmdbuf->buf.fd, &exec);
This might assert, shouldn't it be sth like:
ret = __xe_exec(...);
if (ret)
return ret;
xe_submit_cmd() is asserting, so it might assert from both
reasons - unsuccessful exec or unsuccessful wait.
My personal preference is to not assert in underscore functions
giving caller chance to handle errors. Only exception is to
checking input parameters.
> + ret = __xe_wait_ufence(cmdbuf->buf.fd,
> + (uint64_t *)xe_cmdbuf_exec_ufence_cpuva(cmdbuf),
> + USER_FENCE_VALUE, cmdbuf->exec_queue, &timeout);
> + /* Reset the fence so the exec ufence can be reused */
> + memset((char *)xe_cmdbuf_exec_ufence_cpuva(cmdbuf), 0, UFENCE_LENGTH);
Cosmetic, please add line before return.
> + return ret;
> +}
> +
> +/**
> + * xe_submit_cmd:
> + * @cmdbuf Pointer to the command buffer structure
> + *
> + * Wrapper function to submit a command buffer and assert its successful
> + * execution.
> + */
> +void xe_submit_cmd(struct xe_cmdbuf *cmdbuf)
> +{
> + int64_t ret;
> +
Assert if cmdbuf is not null.
> + ret = __xe_submit_cmd(cmdbuf);
> + igt_assert_eq(ret, 0);
> +}
> +
> +/**
> + *xe_create_buffer:
> + * @buffer Pointer to the xe_buf structure containing buffer details.
> + *
> + * Creates a buffer, maps it to both CPU and GPU address spaces, and binds it
> + * to a virtual memory (VM) space.
> + */
> +void xe_create_buffer(struct xe_buf *buffer)
> +{
> + struct drm_xe_sync sync[1] = {
> + { .type = DRM_XE_SYNC_TYPE_USER_FENCE,
> + .flags = DRM_XE_SYNC_FLAG_SIGNAL,
> + .timeline_value = USER_FENCE_VALUE },
> + };
> +
Ditto.
This function should check if all necessary fields are set in buffer
structure before proceed - fd, vm, size, placement.
Calling this function on not properly initialized buffer likely will
lead to segfault.
> + buffer->bind_queue = xe_bind_exec_queue_create(buffer->fd,
> + buffer->vm, 0);
> + buffer->bind_ufence = aligned_alloc(xe_get_default_alignment(buffer->fd),
> + PAGE_ALIGN_UFENCE);
> + sync->addr = to_user_pointer(buffer->bind_ufence);
> +
> +
Unnecessary double newline.
> + /* create and bind the buffer->bo */
> + buffer->bo = xe_bo_create(buffer->fd, 0, buffer->size,
> + buffer->placement, buffer->flag);
> + buffer->cpu_addr = xe_bo_map(buffer->fd, buffer->bo, buffer->size);
> + xe_vm_bind_async(buffer->fd, buffer->vm, buffer->bind_queue,
> + buffer->bo, 0, buffer->gpu_addr,
> + buffer->size, sync, 1);
> +
> + xe_wait_ufence(buffer->fd, buffer->bind_ufence,
> + USER_FENCE_VALUE, buffer->bind_queue, NSEC_PER_SEC);
> + memset(buffer->bind_ufence, 0, PAGE_ALIGN_UFENCE);
> +}
> +
> +/**
> + * xe_destroy_buffer:
> + * @buffer Pointer to the xe_buf structure containing buffer details
> + *
> + * Destroys a buffer created by xe_create_buffer and releases associated
> + * resources.
> + */
> +void xe_destroy_buffer(struct xe_buf *buffer)
> +{
> + struct drm_xe_sync sync[1] = {
> + { .type = DRM_XE_SYNC_TYPE_USER_FENCE,
> + .flags = DRM_XE_SYNC_FLAG_SIGNAL,
> + .timeline_value = USER_FENCE_VALUE },
> + };
Assert if buffer is null
> + sync->addr = to_user_pointer(buffer->bind_ufence);
> +
> + xe_vm_unbind_async(buffer->fd, buffer->vm, buffer->bind_queue,
> + 0, buffer->gpu_addr, buffer->size, sync, 1);
> + xe_wait_ufence(buffer->fd, buffer->bind_ufence,
> + USER_FENCE_VALUE, buffer->bind_queue, NSEC_PER_SEC);
> + memset(buffer->bind_ufence, 0, PAGE_ALIGN_UFENCE);
> +
> + munmap(buffer->cpu_addr, buffer->size);
> + gem_close(buffer->fd, buffer->bo);
> +
> + free(buffer->bind_ufence);
> + xe_exec_queue_destroy(buffer->fd, buffer->bind_queue);
> +}
> +
> +/**
> + * xe_cmdbuf_insert_store:
> + * @cmdbuf: command buffer where commands will be inserted.
> + * @dst_va Destination virtual address to store the value.
> + * @val Value to be stored.
> + *
> + * Inserts a MI_STORE_DWORD_IMM_GEN4 command into a command buffer, which stores
> + * an immediate value to a given destination virtual address.
> + */
> +void xe_cmdbuf_insert_store(struct xe_cmdbuf *cmdbuf,
> + uint64_t dst_va, uint32_t val)
> +{
> + uint32_t *batch = cmdbuf->buf.cpu_addr;
> +
> + igt_assert(!cmdbuf->closed);
You may check if cmdbuf->write_index + length of store-dword (4 dwords).
As xe_bb_size() adds additional page you won't overwrite memory, just
assert due to lack of space for BBE.
> +
> + batch[cmdbuf->write_index++] = MI_STORE_DWORD_IMM_GEN4;
> + batch[cmdbuf->write_index++] = dst_va;
> + batch[cmdbuf->write_index++] = dst_va >> 32;
> + batch[cmdbuf->write_index++] = val;
> +
> + /* Leaves at least one dword for MI_BATCH_BUFFER_END */
> + igt_assert(cmdbuf->write_index <= cmdbuf->cmd_size/sizeof(uint32_t) - 2);
> +}
> +
> +/**
> + * xe_create_cmdbuf:
> + * @cmdbuf Pointer to the xe_cmdbuf structure representing the command buffer.
> + * @eci Pointer to the engine class instance for execution.
> + *
> + * Creates a command buffer, fills it with commands using the provided fill
> + * function, and sets up the execution queue for submission.
> + */
> +void xe_create_cmdbuf(struct xe_cmdbuf *cmdbuf,
> + struct drm_xe_engine_class_instance *eci)
> +{
> + struct xe_buf *buf = &cmdbuf->buf;
> + /*
> + * make some room for a exec_ufence, which will be used to sync the
> + * submission of this command....
> + */
> + buf->size = xe_bb_size(buf->fd,
> + cmdbuf->cmd_size + PAGE_ALIGN_UFENCE);
> + xe_create_buffer(buf);
> + cmdbuf->exec_queue = xe_exec_queue_create(buf->fd, buf->vm, eci, 0);
> + cmdbuf->write_index = 0;
> + cmdbuf->closed = false;
> +}
> +
> +/**
> + * xe_fill_cmdbuf:
> + * @cmdbuf Pointer to the xe_cmdbuf structure representing the command buffer.
> + * @fill_func Pointer to the function that fills the command buffer.
> + * @dst_gpu_va Destination GPU virtual address to store the value.
> + * @val Value to be stored.
> + *
> + * Fills the command buffer with commands using the provided fill function.
> + */
> +void xe_fill_cmdbuf(struct xe_cmdbuf *cmdbuf,
> + xe_cmdbuf_fill_func_t fill_func,
> + uint64_t dst_gpu_va, uint32_t val)
> +{
> + igt_assert(!cmdbuf->closed);
> + fill_func(cmdbuf, dst_gpu_va, val);
I don't like this function and I vote for it's removal. You should
just operate directly on cmdbuf (it already have everything you need).
xe_cmdbuf_fill_func_t is not generic (limited to gpu_va/val) so what's
for defining it?
So instead:
xe_create_cmdbuf(&cmdbuf, eci);
xe_fill_cmdbuf(&cmdbuf, xe_cmdbuf_insert_store,
dst_buf.gpu_addr, 0xc0ffee);
xe_close_cmdbuf(&cmdbuf);
xe_submit_cmd(&cmdbuf);
I would imagine:
xe_create_cmdbuf(&cmdbuf, eci);
xe_cmdbuf_insert_store(&cmdbuf, dst_buf.gpu_addr, 0xc0ffee);
xe_cmdbuf_insert_bbe(&cmdbuf, ...)
xe_submit_cmd(&cmdbuf);
Any additional commands might be added by generic calls:
xe_cmdbuf_insert_dw(&cmdbuf, 0xc0de);
xe_cmdbuf_insert_dw(&cmdbuf, 0xc0de2);
You may even drop xe_cmdbuf_insert_bbe() by:
xe_cmdbuf_insert_dw(&cmdbuf, MI_BATCH_BUFFER_END);
> +}
> +
> +/**
> + * xe_close_cmdbuf:
> + * @cmdbuf command buffer to be closed.
> + *
> + * After this function is called, no more commands can be added to the
> + * command buffer.
> + */
> +void xe_close_cmdbuf(struct xe_cmdbuf *cmdbuf)
> +{
> + uint32_t *batch = cmdbuf->buf.cpu_addr;
> +
> + igt_assert(!cmdbuf->closed);
> + batch[cmdbuf->write_index++] = MI_BATCH_BUFFER_END;
> + cmdbuf->closed = true;
Please drop this function. You don't need to keep any logic
regarding if buffer has bbe or not.
--
Zbigniew
> + igt_assert(cmdbuf->write_index <= cmdbuf->cmd_size/sizeof(uint32_t) - 1);
> +}
> +
> +/**
> + * xe_destroy_cmdbuf:
> + * @cmdbuf Pointer to the xe_buf structure representing the command buffer.
> + *
> + * Destroys a command buffer created by xe_create_cmdbuf and releases
> + * associated resources.
> + */
> +void xe_destroy_cmdbuf(struct xe_cmdbuf *cmdbuf)
> +{
> + xe_exec_queue_destroy(cmdbuf->buf.fd, cmdbuf->exec_queue);
> + xe_destroy_buffer(&cmdbuf->buf);
> +}
> +
> static bool __region_belongs_to_regions_type(struct drm_xe_mem_region *region,
> uint32_t *mem_regions_type,
> int num_regions)
> diff --git a/lib/xe/xe_util.h b/lib/xe/xe_util.h
> index 06ebd3c2a..26321f957 100644
> --- a/lib/xe/xe_util.h
> +++ b/lib/xe/xe_util.h
> @@ -14,6 +14,53 @@
>
> #include "xe_query.h"
>
> +#define USER_FENCE_VALUE 0xdeadbeefdeadbeefull
> +#define PAGE_ALIGN_UFENCE 4096
> +#define UFENCE_LENGTH 8
> +
> +struct xe_buf {
> + void *cpu_addr;
> + uint64_t gpu_addr;
> + /*the user fence used to vm bind this buffer*/
> + uint64_t *bind_ufence;
> + uint64_t size;
> + uint32_t flag;
> + uint32_t vm;
> + uint32_t bo;
> + uint32_t placement;
> + uint32_t bind_queue;
> + int fd;
> + bool is_userptr;
> +};
> +
> +struct xe_cmdbuf {
> + struct xe_buf buf;
> + /* command size in bytes, not including exec_ufence */
> + uint64_t cmd_size;
> + uint32_t exec_queue;
> + /* Dword index to writ to command buffer */
> + uint32_t write_index;
> + bool closed;
> +};
> +
> +typedef void (*xe_cmdbuf_fill_func_t) (struct xe_cmdbuf *cmdbuf,
> + uint64_t dst_gpu_va, uint32_t val);
> +
> +void xe_create_buffer(struct xe_buf *buffer);
> +void xe_destroy_buffer(struct xe_buf *buffer);
> +
> +void xe_create_cmdbuf(struct xe_cmdbuf *cmdbuf,
> + struct drm_xe_engine_class_instance *eci);
> +void xe_fill_cmdbuf(struct xe_cmdbuf *cmdbuf,
> + xe_cmdbuf_fill_func_t fill_func,
> + uint64_t dst_gpu_va, uint32_t val);
> +void xe_close_cmdbuf(struct xe_cmdbuf *cmdbuf);
> +void xe_cmdbuf_insert_store(struct xe_cmdbuf *cmdbuf, uint64_t dst_va,
> + uint32_t val);
> +void xe_submit_cmd(struct xe_cmdbuf *cmdbuf);
> +int64_t __xe_submit_cmd(struct xe_cmdbuf *cmdbuf);
> +void xe_destroy_cmdbuf(struct xe_cmdbuf *cmdbuf);
> +
> #define XE_IS_SYSMEM_MEMORY_REGION(fd, region) \
> (xe_region_class(fd, region) == DRM_XE_MEM_REGION_CLASS_SYSMEM)
> #define XE_IS_VRAM_MEMORY_REGION(fd, region) \
> --
> 2.26.3
>
More information about the igt-dev
mailing list